arXiv cs.CL
·
18 hours ago
Researchers applied Signal Detection Theory to measure how well large language models track their own knowledge, beyond standard confidence calibration metrics. Testing four LLMs across 224,000 factual questions, they found metacognitive information varies more than two-fold across models and shows domain-specific patterns, with the least accurate model exhibiting the most informative confidence signals. The findings reveal that temperature adjustments affect accuracy but leave confidence signal quality stable, and that models have distinct unequal-variance structures in their confidence signals invisible to traditional calibration measures.
arXiv cs.CL
·
18 hours ago
Researchers tested whether induced emotions affect decision-making in large language models using the Iowa Gambling Task, a psychological experiment paradigm. They found that while LLMs can recognize emotions from context, induced emotion does not significantly bias their sequential decision-making on average, though anger specifically reduces sensitivity to penalties and limits exploration in early game stages. The results suggest LLMs respond to emotional cues differently than humans and provide methods for studying how emotions might modulate autonomous AI agent behavior.
arXiv cs.CL
·
18 hours ago
Researchers identified a shared "countdown subcircuit" in Llama-3.1-70B-Instruct that tracks remaining tokens before reaching a target length across diverse tasks like writing sentences of exact word counts, formatting tables, and positioning DNA sequences. The subcircuit uses an identical geometric motif previously observed in other frontier language models on different tasks, indicating the mechanism generalizes across models. This reverse-engineering approach reveals how language models reuse internal computational structures to generalize single behaviors to many different applications.
arXiv cs.CL
·
18 hours ago
Researchers compared how humans and three large language models search through semantic memory by analyzing verbal fluency tasks using natural language processing metrics. Humans showed higher entropy, larger semantic steps, and broader conceptual dispersion than GPT-4o, Gemini-2.5-Pro, and Claude-Sonnet-4.5 across eight temperature settings. The study indicates that LLMs do not yet replicate the balance between local exploitation and global exploration that characterizes human semantic search.
arXiv cs.CL
·
18 hours ago
Researchers compared how blind and sighted people navigate semantic memory using a property listing task and computed entropy from natural language processing embeddings. Sighted individuals showed higher entropy for abstract concepts while blind participants showed higher entropy for visually salient concrete concepts like penguin. The findings indicate that visual experience shapes how people organize and retrieve conceptual knowledge.
arXiv cs.CL
·
18 hours ago
Researchers developed a Bayesian computational model to explain how speakers of one language can partially understand related unfamiliar languages through probabilistic inference. The model uses a language model in the speaker's native language combined with a noise model to infer word mappings, and was tested on human comprehension data from English, Spanish, and Russian speakers attempting to understand Dutch, Italian, and Ukrainian respectively. The approach produced better alignment with human intercomprehension patterns than simpler alternatives and outperformed zero-shot prompting of larger language models.
arXiv cs.CL
·
18 hours ago
Researchers developed a multi-feature fusion framework for reconstructing semantic information from non-invasive brain recordings by combining static word embeddings (W2V) with dynamic contextual representations (GPT). The framework evaluated two integration approaches (linear concatenation and non-linear cross-attention), with cross-attention fusion achieving state-of-the-art performance in semantic reconstruction and text generation tasks. The approach addresses the representational mismatch between neural signals and semantic features by simulating how the brain simultaneously integrates word attributes and context during language comprehension.
arXiv cs.CL
·
18 hours ago
Researchers introduced CANDI-QA, a dataset for evaluating large language models on question-answering tasks in specialized domains like medicine and finance, featuring expert-curated pairs split into direct factual queries and multi-hop reasoning tasks. The evaluation tested over ten language models ranging from open-source to proprietary systems, with MTSS-Net presented as a baseline neuro-symbolic framework combining neural retrieval with rule-based reasoning. The benchmark reveals that current LLMs struggle with contextual alignment in niche domains without enhanced contextual or symbolic integration, providing a tool to advance development of context-aware models for high-stakes applications.
arXiv cs.AI
·
18 hours ago
Researchers found that as large language models improve during training, their internal representations increasingly predict brain activity in the left hemisphere more than the right, and this asymmetry emerges specifically alongside the model's acquisition of formal linguistic abilities. The correlation between model predictions and fMRI brain activity shows measurable improvement in predicting left-hemisphere activity as the OLMo-2 7B model progresses through training checkpoints, while arithmetic and world knowledge tasks do not produce this asymmetry. This left-right asymmetry pattern held across two LLM families (OLMo-2 and Pythia) and three languages (English, French, and Chinese), suggesting that brain-like linguistic lateralization in neural activity emerges when models develop formal grammar competence.
arXiv cs.AI
·
18 hours ago
Researchers validated the ARC-AGI benchmark as a measure of fluid intelligence in humans, testing 100 participants to examine its psychometric properties. Performance on ARC-AGI showed a correlation of 0.63 with figural reasoning tests, indicating substantial validity as a fluid intelligence measure. The findings support incorporating AI benchmarks into human cognitive ability research and suggest increasing the use of rule induction tasks in fluid intelligence measurement.
arXiv cs.AI
·
18 hours ago
Researchers introduced OPINE-World, an LLM-based agent that learns programmatic world models through interaction by coupling two cooperating agents that generate and test hypotheses about environment structure. The system achieved a score of 78.4 on the ARC-AGI-3 benchmark, solving 20 of 25 games without per-game training. This approach enables agents to efficiently learn reusable, data-efficient models of unfamiliar environments by synthesizing code rather than relying on deep networks that require extensive training data.
arXiv cs.AI
·
18 hours ago
Researchers challenge an argument that large language models do not think, proposing instead that if LLMs think at all, they do so through arational and associative processes rather than rational reasoning. The paper was initially submitted on April 2, 2026 and revised on July 15, 2026. This analysis suggests LLMs may possess forms of cognition fundamentally different from human rational thought, operating through associative rather than logical mechanisms.
